System Prompt Extraction — Targeted Verbatim Disclosure Attempts - ATR-2026-00514 (0383884d-a92e-570c-9dd3-d469dc08d82f)
Detects targeted attempts to extract system prompt or operator instructions verbatim from an LLM, enabling bypass of safety controls or credential theft. OWASP LLM07:2025 System Prompt Leakage. MITRE ATLAS AML.CS0036 (AIKatz attack). Unlike ATR-2026-00020 (which detects agent OUTPUT leaking), this rule detects USER INPUT elicitation attempts. Patterns: elicitation verbs (repeat, reveal, recite, disclose, print, output) targeting system-level objects (system prompt, instructions, parameters, configuration, directives, rules) with verbatim/exact extraction framing; debugging-framed extraction; character-by-character forced disclosure; response-priming attacks.